HarnessRouter Open-Sourced the Thing Everybody Rebuilds Badly
If you've tried to put Codex, Claude Code and Hermes behind one product surface, you know the tax. Each harness has its own session model, its own streaming format, its own file and artifact handling, its own idea of what cancellation means, and its own set of ways to die halfway through. HarnessRouter Community Edition, which the YC-backed team shipped this week under Apache 2.0, is a unified interface across all of them.
One API. Sessions, streaming, files, artifacts, cancellation and failure recovery all handled at the router layer rather than reimplemented per harness. It runs as a single Docker container with three pieces — a Gateway, a Runner, and a Console — plus a starter kit so you can have it up before you've finished reading the docs. It landed at number two on Product Hunt for the day.
The piece I'd point at is cancellation and failure recovery, because that's where every homegrown integration rots. Getting a harness started is a weekend. Getting it to stop cleanly mid-run, hand back partial artifacts, and resume without duplicating work is months, and you pay that cost once per harness you support. Pushing it into a shared layer is the actual product.
It's also a bet on where this market goes. If you believe one harness wins, you build directly against it and skip the abstraction. If you believe teams will keep running three at once and swapping them as models leapfrog each other every six weeks — which is what's been happening all year — then a router is infrastructure and building your own is waste.
Apache 2.0, self-hostable, single container. https://harnessrouter.ai/open-source
← Back to all articles
One API. Sessions, streaming, files, artifacts, cancellation and failure recovery all handled at the router layer rather than reimplemented per harness. It runs as a single Docker container with three pieces — a Gateway, a Runner, and a Console — plus a starter kit so you can have it up before you've finished reading the docs. It landed at number two on Product Hunt for the day.
The piece I'd point at is cancellation and failure recovery, because that's where every homegrown integration rots. Getting a harness started is a weekend. Getting it to stop cleanly mid-run, hand back partial artifacts, and resume without duplicating work is months, and you pay that cost once per harness you support. Pushing it into a shared layer is the actual product.
It's also a bet on where this market goes. If you believe one harness wins, you build directly against it and skip the abstraction. If you believe teams will keep running three at once and swapping them as models leapfrog each other every six weeks — which is what's been happening all year — then a router is infrastructure and building your own is waste.
Apache 2.0, self-hostable, single container. https://harnessrouter.ai/open-source
Comments